Episode IX of Phase One: Orchestrating docs, AI accuracy, and the pipeline writer
Ninth episode of the AI & Docs podcast series is up! In this one, Tom and I talk with Sarah Deaton, a technical writer at Anthropic, about what documentation work looks like when automation handles most content generation: shifting from writing to orchestrating, mining AI assistant conversations for gaps, building forensic processes for accuracy claims, and managing high-volume PR pipelines without losing quality.
You can watch / listen to the episode here:
Some of the things I said:
On documentation observability
The dream is to have some sort of documentation observability, fully anonymized, where you see maybe the journey of a user through product and docs alike.
On model preferences for development work
I do lots of development work these days, more than writing, for docs infrastructure and stuff like that. And when I use Claude, I use plan mode a lot for complex stuff, and I rely on complex models, heavy models, for the planning part.
On skills and reference content
There’s always a significant overlap between a product skill and the docs…the reference bits are the ones that I’d rather have as links to the docs.
On Simplified Technical English
You just put in the base prompt of Claude or any AI, ‘always write using Simplified Technical English,’ and it works wonders. It’s way better than the caveman style.
On security models for agentic workflows
I really like the approach that GitHub is following with agentic workflows, because they have a very strong security model where essentially they run the agentic session in a flow where there’s a firewall that checks the networking.